DETU-Net: An Improved TransUNet with Dual-Branch for Right Ventricle MRI Segmentation

Cardiovascular diseases related to the right ventricle have a high mortality and disability rate worldwide. Accurately segmenting the right ventricle from cardiac magnetic resonance imaging is essential for the diagnosis and analysis of cardiac diseases. However, owing to the intricate crescentshaped geometry of the structure, blurry boundary features, and significant shape variation from the apex to the base, automatic segmentation of the RV still faces tremendous challenges. Existing methods still lack accuracy in dealing with the complexity and deformability of such structures. Therefore, we propose a DETUNet model to segment right ventrilce, which mainly consists of two components: the dual-branch module and the ExTransformer module. The Dual-branch module employs a parallel branch structure to capture detailed information, thus enabling the extraction of the edge features of RV. ExTransformer enhances the long-range dependencies of the model, not only to focus on its own global features, but also to enhance the modeling of RV regions with structural discontinuities or large morphological variations in the form of cross-slices. Comparative experiments are conducted on both RVSC and ACDC datasets. The results on the RVSC dataset show that DETU-Net achieved a mean Dice on the epicardium and endocardium of 0.88 and 0.91, which demonstrates that our segmentation approach outperforms representative approaches.